Vercel's $1M sandbox challenge finds a kernel zero-day reaching the host

Vercel's sandbox escape challenge drew 1,285 reports and paid $242,000 in bounties. The top find was a Linux kernel networking flaw that let a microVM guest leak host memory, crash the host or run code, with no customer data accessed. Read: Vercel's sandbox escape challenge drew 1,285 reports and paid $242,000 in bounties. The top find was a Linux kernel networking flaw that let a microVM guest leak host memory, crash the host or run code, with no customer data accessed. Read: Anthropic now exports OpenTelemetry traces from the Messages API, Managed Agents sessions and the Claude SDKs, and LangSmith accepts them on day one. Read: SemiAnalysis says Google submitted TPUv7 FP8 vLLM results to its Agentic Inference Benchmark, showing 1.67x better performance per dollar than GB300 NVL72 at FP4. Read: Subframe's Real Design Benchmark asks people to choose between a real product screen and one built by Opus 5.5 or GPT-6.1 Sol. After thousands of votes, 57% preferred the AI design. Read: Epoch AI's SpireClimb has models play Slay the Spire through the game's graphical interface. Frontier models now win over 90% of easy runs and 20% at the hardest difficulty. Read: Siemens EDA used NVIDIA NeMo Gym verifiers to find failure modes in a chip-design agent, then rewrote its skills and MCP tool descriptions. Its benchmark score rose from about 55% to 95% as turns fell from 155 to 53. Read: SWEeper-Bench gives coding agents a codebase and an open-ended request, so they must write interactive tests to find bugs in 200 real web apps. The best of 15 agents passes 59%.